Graph contrastive learning alleviates the data sparsity issue in recommendation systems by exploring diverse representations of the same node. Most existing contrastive learning methods generate contrastive views by either pruning interaction graphs or introducing perturbations in the latent space. However, these approaches fail to effectively address the cold-user problem. Previous studies have attempted to mitigate data sparsity and cold-user challenges by leveraging exogenous data such as social networks. Nonetheless, these methods suffer from inherent group limitations within social networks. This paper proposes a graph contrastive learning model based on endogenous group relationship networks (EgrGCL). The model employs a graph convolutional network to capture high-order collaborative signals from both the bipartite interaction graph and the endogenous group relationship graph. Additionally, a simplified dual-channel contrastive learning framework is introduced to learn distinct representations of the same user node in both the personalized preference space and the endogenous group relationship space, thereby enhancing the quality of node embeddings for recommendation tasks. Extensive experiments on three real-world datasets demonstrate the superiority of EgrGCL over baseline methods, particularly in addressing the cold-user problem.